FastH3’s 8-step audio-video model gets a ComfyUI-ready release

FastVideo has repackaged its distilled MiniMax H3 checkpoint for ComfyUI. The new FastH3 Comfy repo provides the FastH3 8-Step V2 weights as ComfyUI-compatible safetensors files, allowing local text-to-video generation with synchronized audio without manual checkpoint conversion or the full FastVideo serving stack.

At the time of publication, the repository had nearly 11,000 downloads, and ComfyUI’s official MiniMax H3 workflow listed it as a supported checkpoint. The release simplifies model loading, although developers still need substantial GPU memory and a VSA-H3-compatible attention implementation.

Eight passes replace 49

FastH3 is a distilled version of MiniMax H3, an open-weight model that generates video and audio together. The base model’s standard schedule evaluates its 33-billion-parameter audio-video diffusion transformer 49 times for each generation. FastH3 reduces that workload through fewer evaluations and sparse attention, which limits how many token relationships the transformer computes.

The ComfyUI release contains the 8-step V2 checkpoint. It uses Distribution Matching Distillation, or DMD2, to train a smaller sampling schedule that approximates the base model’s output distribution in eight transformer evaluations. The checkpoint retains H3’s Qwen3-VL text encoder, video VAE, and audio VAE, preserving the same inputs and output types while reducing inference work.

The repository identifies this checkpoint as the step-1300 model, meaning it was saved after 1,300 distillation training steps. Its data-free DMD2 process uses teacher-generated samples in place of the original training corpus. VSA-H3 applies 80% attention sparsity, pruning roughly four-fifths of the attention calculations. The required video scheduler shift is 10; the base H3 value of 12 does not match this distilled checkpoint.